acceptodds
Under review as a conference paper at ICLR 2027

Bridging Regular and Irregular Time Series: A Unified Perspective

Abstract

Most time-series forecasting models assume that all variables are observed on a shared grid with uniform temporal spacing. Real-world systems often violate this assumption: in irregular multivariate time series (IMTS), variables are observed asynchronously at non-uniform timestamps, with observation frequencies varying across variables. Canonical pre-alignment (CPA) maps these observations to a shared time–variable grid, thereby placing regular and irregular time series in a common representation. This representation raises a fundamental question: are forecasting models developed for regular time series truly incapable of handling irregular observations? To address this question, we establish a unified perspective on representative models developed for irregular time series by organizing them according to a common four-stage pipeline: normalization, input embedding, dependency modeling, and forecast generation. By combining this perspective with controlled experiments, we identify the key factors for effective irregular forecasting and develop BridgeTST from these findings. Across five irregular and five regular datasets, BridgeTST achieves strong forecasting performance in both settings. In contrast, representative irregular forecasting models generally perform poorly in the regular setting. Building on these findings, we further examine the boundary between regular and irregular forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_KiptOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper examines the relationship between regular and irregular time series forecasting. It organizes representative methods into four stages—normalization, input embedding, dependency modeling, and forecast generation—and develops BridgeTST through comparisons at each stage. The model combines observed values, masks, timestamps, and variable identities on a shared timeline, using general-purpose embedding, attention, and forecasting components. Experiments cover both regular and irregular data, with module replacements used to examine the role of specialized designs.

Critical strengths
  1. The research question is valuable. The paper distinguishes backbone capabilities from the roles of input information and the processing pipeline. This helps reassess why models designed for regular data underperform on irregular data and offers a useful direction for building stronger baselines.
  2. The analysis has a reasonably clear empirical basis. The four-stage framework clarifies the common structure of different methods. Comparisons of normalization, input information, cross-variable interactions, and prediction heads support the selected design choices.
  3. The method has practical value in both settings. BridgeTST is competitive on the evaluated regular and irregular datasets, supporting its architecture built from general-purpose components as an important reference for irregular forecasting.
Critical weaknesses
  1. The handling of future target timestamps needs clarification. The appendix specifies each dataset's alignment resolution and forecast horizon, but does not make clear whether these settings allow target timestamps to be inferred from output positions. Equation (3) conditions forecasts on future query timestamps, whereas Equation (8) and the architecture diagram do not show how they are used. Clarifying this correspondence would establish which query settings the fixed-length prediction head supports.
  2. Direct evidence for why existing regular forecasting models underperform remains limited. The stage-by-stage analysis supports BridgeTST's effectiveness, but the paper further suggests that the limitations of regular forecasting models may stem from normalization and input information rather than their backbones. The main experiments do not sufficiently explain how the regular baselines were adapted in these respects, or clearly demonstrate the effects of such adaptations on representative existing backbones. The current evidence more directly supports the claim that general-purpose components can form an effective model; it remains insufficient to establish the main causes of the performance gaps of existing regular models.
Questions for authors
  1. Do the alignment and forecasting settings in the appendix mean that the future target timestamps in Equation (3) are implicitly determined by output positions and need not be supplied explicitly? Please clarify how this relates to Equation (8). If target timestamps are not fixed, how are outputs matched to query timestamps?
  2. The paper suggests that the limitations of existing regular forecasting models on irregular data may arise from normalization and the handling of sampling information. Do the experiments support this explanation, or only establish that these design choices are effective within BridgeTST? Please clarify the scope of the conclusion accordingly.
  3. Figure 3 shows that the module replacements degrade performance, but does not establish the specific causes of that degradation. The proposed mechanistic explanation should therefore be qualified.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Kipt (Part I)

Thank you for your detailed comments on unclear equations, evidence for adapting regular models, and the explanation of module-replacement effects.

W1&Q1: The use of future target timestamps needs clarification.

A:

Equation (3)‾\underline{\text{Equation (3)}} uses the set of future target timestamps Q\mathbf{Q} as a prediction condition. The MLP head in Equation (9)‾\underline{\text{Equation (9)}} always outputs FF ordered predictions and does not explicitly read Q\mathbf{Q}. Under our fixed-window forecasting protocol, the ff-th output corresponds to the ff-th target time step. Thus, the head supports a fixed, ordered forecast window. It does not support queries at arbitrary timestamps.

We compare the FTC Head with and without future timestamps. The results are shown below.

DatasetWith Future TimestampsWithout Future TimestampsMSE change (%)MAE change (%)
MetricMSE / MAEMSE / MAE——
HumanActivity0.0442 / 0.12090.0445 / 0.1212+0.7%+0.2%
USHCN0.1494 / 0.20610.1489 / 0.2060-0.3%-0.0%
P120.3091 / 0.35190.3095 / 0.3516+0.1%-0.1%
MIMIC-III0.4056 / 0.36300.4015 / 0.3605-1.0%-0.7%
MIMIC-IV0.2444 / 0.28600.2433 / 0.2860-0.5%+0.0%

Across the five irregular datasets, future timestamps have only a small effect on performance. In the settings studied here, explicitly providing future timestamps makes little difference.

W2&Q2: Direct evidence about why regular forecasting models perform poorly is limited.

A: To further study this performance gap, we conduct adaptation experiments on PatchTST.

Model and settingHumanActivityUSHCNP12MIMIC-IIIMIMIC-IV
MetricMSE / MAEMSE / MAEMSE / MAEMSE / MAEMSE / MAE
PatchTST: RevIN0.0725 / 0.17860.1681 / 0.25580.3276 / 0.36990.5327 / 0.42210.3323 / 0.3605
PatchTST: Masked RevIN0.0481 / 0.12480.2071 / 0.27880.3357 / 0.37010.5755 / 0.44180.3353 / 0.3409
PatchTST: no RevIN0.0606 / 0.16400.1558 / 0.24530.3277 / 0.37200.4824 / 0.41760.3574 / 0.3945
AuthorsOpenReview ↗

Response to Reviewer Kipt (Part II)

W2&Q2 (continued: input-feature ablation).

Model and settingHumanActivityUSHCNP12MIMIC-IIIMIMIC-IV
MetricMSE / MAEMSE / MAEMSE / MAEMSE / MAEMSE / MAE
PatchTST: no RevIN, values only0.0606 / 0.16400.1558 / 0.24530.3277 / 0.37200.4824 / 0.41760.3574 / 0.3945
PatchTST: no RevIN, X+M\mathcal X+\mathcal M0.0547 / 0.15100.1775 / 0.28680.3169 / 0.36490.4648 / 0.40570.2891 / 0.3291
PatchTST: no RevIN, X+M+E\mathcal X+\mathcal M+\mathcal E0.0667 / 0.18090.1579 / 0.23450.3161 / 0.36250.4479 / 0.39230.2678 / 0.3098
PatchTST: no RevIN, X+M+E+T\mathcal X+\mathcal M+\mathcal E+\mathcal T0.0668 / 0.17620.1543 / 0.23020.3157 / 0.36130.4408 / 0.38680.2611 / 0.3035

Overall, these results lead to conclusions broadly consistent with those drawn from BridgeTST, with some differences across individual datasets. Special design choices in regular models may introduce additional factors. For example, PatchTST processes each channel independently. This may lead to much worse performance on irregular data than irregular-specific models and obscure the underlying cause. BridgeTST starts from a standard patch-based architecture commonly used for regular forecasting. This makes it a more controlled starting point than regular models with diverse architectures.

The PatchTST normalization and input ablations are reported in Appendix A.7.1, Tables 13–14‾\underline{\text{Appendix A.7.1, Tables 13–14}}.

Q3: The module-replacement results show a performance drop but do not identify its cause. The paper should therefore limit its mechanistic explanation.

A: To better understand the performance degradation observed in Figure 3, we further conduct ablation studies on TTCN.

TTCN forms a patch representation by weighting each position's input with a position-specific filter coefficient, applying the same operation across output channels. For one output channel, let tt index a patch position; its input combines the observed value and timestamp encoding:

ut=[mtxt;ϕ(τt)].\mathbf u_t=[m_tx_t;\phi(\tau_t)].

Here, xtx_t is the observed value, mtm_t indicates whether the position is valid, and ϕ(τt)\phi(\tau_t) is the timestamp encoding. The original TTCN computes coefficients from the current input. We also introduce a static ablation in which the coefficients do not depend on the input. They are learned directly:

g_\theta(\mathbf u_t), & \text{TTCN},\\\\ \mathbf s_t, & \text{static TTCN}. \end{cases}$$ We also study how TTCN aggregates observations using these coefficients. The original TTCN uses softmax. We compare it with raw sum and mean sum, where $\mathcal V$ denotes the valid time positions: $$\mathbf w_t=\begin{cases} \operatorname{softmax}_{t\in\mathcal V}(\mathbf a_t), & \text{softmax},\\\\ \mathbf a_t, & \text{raw sum},\\\\ \mathbf a_t/\max(1,|\mathcal V|), & \text{mean sum}. \end{cases}$$ Softmax normalizes coefficients over valid positions. Raw sum uses the coefficients directly. Mean sum divides the sum by the number of valid positions. The weighted inputs are then summed to produce the output for that channel: $$z=\sum_{t\in\mathcal V}\mathbf w_t^\top\mathbf u_t.$$ We replace the input embedding in BridgeTST and compare five settings on ETTh1 and ETTm1 with an input length of 512 and a forecast horizon of 96: MLP, Original TTCN, Static TTCN with softmax, Static TTCN with raw sum, and Static TTCN with mean sum. | Input embedding | ETTh1 | ETTm1 | |---|---:|---:| | **Metric** | **MSE / MAE** | **MSE / MAE** | | MLP | **0.3650 / 0.3892** | **0.2818 / 0.3281** | | Original TTCN | 0.6830 / 0.5550 | 0.3745 / 0.4135 | | Static TTCN / softmax | 0.6504 / 0.5737 | 0.3332 / 0.3811 | | Static / raw sum | 0.4441 / 0.4585 | 0.3040 / 0.3488 | | Static / mean sum | 0.5974 / 0.5441 | 0.3442 / 0.3837 | Replacing softmax with raw sum substantially reduces the error, but the result is still worse than MLP. This suggests that softmax aggregation accounts for part of TTCN's performance drop. We have limited the mechanistic claims in the revised manuscript accordingly. The TTCN ablation is described in $\underline{\text{Section 5.3.2}}$, with results in $\underline{\text{Table 7}}$.
Reviewer_KiptOpenReview ↗

Official Comment by Reviewer Kipt

Thank you for the clarifications and additional experiments, which substantially address my main concerns about the forecasting setup and the evidence supporting the paper’s conclusions. I have no further questions and keep my score.

AuthorsOpenReview ↗

Thank You for Your Feedback

Thank you for reviewing our clarifications and additional experiments. We appreciate your feedback and are glad that our response addressed your concerns.

Reviewer_zYgHOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether general-purpose forecasting components can handle irregular multivariate series on a shared grid of values, timestamps and observation masks. It organises five irregular forecasters into normalization, embedding, dependency modeling and forecasting and derives BridgeTST as their best combination. BridgeTST performs best in an evaluation against several regular and irregular forecasting baselines on five datasets.

Critical strengths
  • BridgeTST performs best in an evaluation against several regular and irregular forecasting baselines on five datasets.
  • Existing approaches are well oragnised into a common pipeline.
Critical weaknesses
  • It is not very clear if the paper's focus is on the organisation/description of irregular forecasting models and their analysis in comparison with regular models, or on the provision of a strong new model. In both aspects, the paper could do stronger: for analysis, the evaluation and analysis is rather limited, for a new model BridgeTST is just a combination of existing models. Details see next comments.
  • There is no dedicated related work section and missing information regarding the selection of the five "representative irregular models".
  • The purpose of Section 3 is not fully clear. Is it related work, is it about explaining the steps (in that case, I would expect more formal details about normalisation, for example) or to explain the creation of BridgeTST?
  • The main finding, already introduced to the central research question "Are regular models truly incapable of handling irregular observations?", is that "a model developed for regularly sampled data can also handle irregular observations once we identify the choices that matter for irregular observations". But these choices are exactly what distinguish irregular from regular models.
  • The best combination of components is done as follows: "At each stage, we vary one component while holding the others fixed and carry the selected configuration forward". This leads to the question if there are any unseen dependencies between the components whose combinations are not tested.
  • GraFITi and HyperIMTS, the only irregular baselines with neither TTCN nor an FTC head (Table 1) are not tested on regular datasets. Therefore, the claim that irregular models underperform on regular data is limited to the three FTC-head models
  • Figure 3 omits Electricity and Traffic, where the difference to the irregular models is much bigger than with the other three datasets.
  • The claim "Yet the model performs better without normalization in most cases. These results expose an overlooked issue in prior research: regular models are often evaluated on irregular data with RevIN still enabled" (Section 4.1) indicates that regular models can look unable to handle irregular data when RevIN is the real problem. However, this claim is not analysed further by testing different normalisation models on the regular models. There is also no information given what normalisation is used by the regular models.
  • No code is provided for review.

Minor comments:

  • For (human+machine) readability, better provide figures as embedded PDFs/SVGs.
  • Section 4 repeatedly refers to Section 3, but should actually refer to proper subsections (3.1 etc)
  • Figure 3 should describe in the caption regarding what model the change is reported, and the figure itself seems a bit over-styled (e.g., non-standard font).
  • Conclusion should be more self-contained (first sentence: models for what?).
  • The reproducibility statement refers to the non-existing Section 7.
Questions for authors
  • Q1: Is there any evidence that the sequential identification of the best components is sufficient? Could there be untested combinations of components that perform better?
  • Q2: Which normalisation strategies are used by the regular models? If they suffer from normalisation, can this be seen in experimental results?
  • Q3: Are there more irregular models that cannot be categorised into the proposed pipeline? If yes, why not?
3 comments
AuthorsOpenReview ↗

Response to Reviewer zYgH (Part I)

Thank you for your detailed comments on the paper's research focus, four-stage analysis, component interactions, and evaluation coverage.

W1&W3: The paper's research focus, analysis scope, and the novelty of BridgeTST need further clarification.

A: Our central question is whether models designed for regular time series can handle irregular observations. Section 3‾\underline{\text{Section 3}} reviews related methods and analyzes their designs through a four-stage framework. Section 4‾\underline{\text{Section 4}} then uses controlled experiments to build BridgeTST step by step. Together, these sections address our research question.

The revised manuscript now includes a discussion of related work and explains why we selected the five models for analysis. It also provides more detail, including formulas for the statistics used by RevIN and Masked RevIN, and descriptions of TTCN, TAPA, and the FTC Head.

Regarding the concern that BridgeTST is only a combination of existing models, the experiments in Section 4‾\underline{\text{Section 4}} guided its design. We did not aim to add complex components. BridgeTST provides a simple and effective baseline for irregular forecasting. The analysis behind its construction may also help researchers design irregular-specific models.

W2: The paper lacks a separate related-work section and does not explain why these five irregular models were selected as representative methods.

A: The revised manuscript now includes a discussion of related work and the rationale for selecting the five models at the beginning of Section 3‾\underline{\text{Section 3}}.

W4: The distinction between irregular and regular models

A: Prior work on irregular forecasting often distinguishes irregular models from regular models by using specialized modules for irregular observations. For example, tPatchGNN uses TTCN, while APN uses TAPA to aggregate irregular observations. Methods that do not use canonical pre-alignment (CPA), such as GraFITi and HyperIMTS, model dependencies with specialized structures. GraFITi uses a bipartite graph, while HyperIMTS uses a hypergraph.

Our finding points to a broader design principle: with suitable input information, basic modeling components can also perform well without specialized modules.

W5&Q1: Could there be undiscovered dependencies between components?

A: To evaluate more component combinations and test for possible interactions, we evaluate two additional configurations: TTCN + Attention + FTC and TTCN + GNN + FTC. We compare both with the BridgeTST baseline.

DatasetBridgeTSTTTCN + Attention + FTCTTCN + GNN + FTC
MetricMSE / MAEMSE / MAEMSE / MAE
HumanActivity0.0413 / 0.11770.0507 / 0.13450.0528 / 0.1362
USHCN0.1408 / 0.21170.1366 / 0.19600.1390 / 0.1964
P120.3056 / 0.35190.3105 / 0.35600.3145 / 0.3568
MIMIC-III0.3910 / 0.36080.4070 / 0.36440.4190 / 0.3700
MIMIC-IV0.2326 / 0.28270.2508 / 0.29100.2489 / 0.2848

Combining TTCN and the FTC Head improves both metrics on USHCN, but does not provide consistent gains on the other datasets. The interaction ablation and its results are reported in Appendix A.7.2, Table 15‾\underline{\text{Appendix A.7.2, Table 15}}.

W6: GraFITi and HyperIMTS have not been evaluated on regular datasets.

A: We evaluate both models on ETTh1 and ETTm1 with an input length of 512 and a forecast horizon of 96.

DatasetModelTest resultRelative increase vs. BridgeTST
MetricMSE / MAEMSE / MAE
ETTh1BridgeTST0.3650 / 0.3892—
ETTh1GraFITi0.7587 / 0.6249+107.9% / +60.6%
ETTh1HyperIMTS0.8290 / 0.6576+127.1% / +69.0%
ETTm1BridgeTST0.2818 / 0.3281—
ETTm1GraFITi0.6489 / 0.5851+130.3% / +78.3%
ETTm1HyperIMTS0.7280 / 0.5931+158.4% / +80.8%

On both datasets, GraFITi and HyperIMTS have substantially higher forecast errors than BridgeTST. Both methods use graphs or hypergraphs over individual observations, which are computationally costly for dense regular time series. Limited compute and rebuttal time prevented us from evaluating them on the remaining regular datasets.

AuthorsOpenReview ↗

Response to Reviewer zYgH (Part II)

W7: Figure 3 does not include results for Electricity and Traffic.

A: We report additional module-replacement results for Electricity (ECL) and Traffic. The table reports the average MSE and MAE over four forecast horizons.

DatasetBridgeTSTTTCNFTC HeadTTCN + FTC Head
MetricMSE / MAEMSE / MAEMSE / MAEMSE / MAE
Electricity0.1561 / 0.24830.1913 / 0.29000.8356 / 0.69180.7894 / 0.6678
Traffic0.4032 / 0.24710.6423 / 0.32491.2798 / 0.68731.2040 / 0.6329

Replacing one or both modules increases the error on both ECL and Traffic, matching the trend on the other three datasets. Figure 3 in Section 5.3.1‾\underline{\text{Figure 3 in Section 5.3.1}} reports the results for all five regular datasets.

W8&Q2: The normalization methods used by regular baselines and their effects have not been adequately described or evaluated.

A: The table below lists the normalization methods used by the regular models.

ModelNormalization method
DLinearNo explicit normalization
PatchTSTRevIN
CrossformerNo explicit normalization
iTransformerRevIN
TimeMixerNormalize each scale independently

We also evaluate different normalization methods on the regular forecasting model PatchTST. The results are shown below.

Model and settingHumanActivityUSHCNP12MIMIC-IIIMIMIC-IV
MetricMSE / MAEMSE / MAEMSE / MAEMSE / MAEMSE / MAE
PatchTST: RevIN0.0725 / 0.17860.1681 / 0.25580.3276 / 0.36990.5327 / 0.42210.3323 / 0.3605
PatchTST: Masked RevIN0.0481 / 0.12480.2071 / 0.27880.3357 / 0.37010.5755 / 0.44180.3353 / 0.3409
PatchTST: no RevIN0.0606 / 0.16400.1558 / 0.24530.3277 / 0.37200.4824 / 0.41760.3574 / 0.3945

Overall, these results lead to conclusions broadly consistent with those drawn from BridgeTST, with some differences across individual datasets. RevIN is harmful in most cases, and Masked RevIN does not fully resolve this issue. One possible reason is that special design choices in regular models introduce additional factors. For example, PatchTST processes each channel independently. This design may lead to much worse performance on irregular data than irregular-specific models, which can obscure the underlying cause. BridgeTST starts from a standard patch-based architecture commonly used for regular forecasting. This makes it a more controlled starting point than regular models with diverse architectures.

The PatchTST normalization ablation is reported in Appendix A.7.1, Table 13‾\underline{\text{Appendix A.7.1, Table 13}}.

W9: Code

The code is included in the supplementary material‾\underline{\text{supplementary material}}.

Q3: The paper does not explain whether the four-stage framework can cover other irregular models.

A: The proposed four-stage framework is intended to provide a general analytical perspective beyond the five representative irregular models examined in this paper. For example, GRU-D, which is not included in our analysis, can also be mapped to this framework. It has no explicit normalization layer. Its input uses observed values, missingness masks, and time intervals, and it handles missing values through temporal decay. Its dependency modeling uses GRU state updates with temporal decay. For forecasting, a readout layer can then generate future values. This example illustrates that the framework can also accommodate models with different architectures.

Minor Comments

  • C1: Figures 1–3‾\underline{\text{Figures 1–3}} are now embedded in the revised manuscript as vector PDFs.
  • C2: References from Section 4‾\underline{\text{Section 4}} to Section 3‾\underline{\text{Section 3}} now point to the relevant subsections (3.1–3.4)‾\underline{\text{subsections (3.1–3.4)}}.
  • C3: Figure 3‾\underline{\text{Figure 3}} is now a single-row panel plot covering five datasets. Its fonts are consistent, and the caption explains the module replacements and error increases.
  • C4: The first sentence of the conclusion‾\underline{\text{conclusion}} now clearly states the research subject.
  • C5: The reproducibility statement now points to the appendix‾\underline{\text{appendix}} instead of the nonexistent Section 7‾\underline{\text{Section 7}}.
AuthorsOpenReview ↗

Follow-up on Our Response

With seven hours remaining in the discussion period, we wanted to kindly follow up and ask whether our response addresses your main concerns. If so, we would be grateful if you would consider revisiting your score. We would also welcome any further feedback.

Reviewer_cd58OpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
3
Soundness
3(+1)
Read review
Paper summary

Addressing the question of whether irregular time series forecasting requires specialized, complex architectures—and whether regular and irregular time series can be handled within a unified approach—this paper proposes a four-stage unified analysis framework (normalization, input embedding, dependency modeling, and prediction generation).

Critical strengths

1.A unified analytical perspective provides significant value: the paper consolidates models with vastly different architectures into a standardized four-stage pipeline—normalization, input embedding, dependency modeling, and prediction generation. Furthermore, it interprets graph/hypergraph operations as sparse attention within a CPA (Coordinate-Position-Attribute) framework, which enhances both comparability and interpretability.

2.The controlled experimental design is clear: by replacing individual components to isolate key factors, the authors conclude that "general-purpose components, with minor adjustments, are sufficient to handle irregular observations." Based on this, they propose BridgeTST—a method that is both simple and highly reproducible.

3.The validation is comprehensive and insightful: the authors benchmark BridgeTST against a wide range of baselines across five irregular and five regular datasets, demonstrating its competitiveness in both settings. Additionally, by conducting experiments that replace specific modules with TTCN/FTC heads, they explain why components designed for irregular data underperform on regular data, thereby establishing a robust and complete line of reasoning.

Critical weaknesses

1.Insufficient evidence for the core claim: The paper asserts that "irregular-pattern forecasting does not require specialized architectures, as general-purpose components suffice." However, the controlled experiments rely primarily on incremental hyperparameter tuning of a single standard patch-based model. The evaluation is limited to only 5+5 datasets, with module-replacement experiments conducted on just three datasets featuring irregular patterns. Furthermore, there is a lack of consistent hyperparameter tuning, loss functions, and computational budgets across all baselines. This conclusion is overly strong and undermines the paper's primary contribution.

2.Questionable fairness in comparisons: BridgeTST employs an L1 loss, whereas the baselines retain their original training objectives. Additionally, the use of RevIN is toggled manually based on whether the data is "regular" or "irregular," and some baselines for irregular patterns are marked as "N/A" on regular datasets due to computational infeasibility.

3.Lack of robustness in the "unified framework" and its underlying mechanisms: BridgeTST still relies on switching the RevIN configuration based on data type, which falls short of a truly unified framework. The explanation for why specialized modules lead to performance degradation on regular data—based primarily on TTCN/FTC head-replacement experiments—lacks granular diagnostic insights or theoretical analysis. Moreover, the four-stage conceptual framework is an oversimplification when applied to non-CPA (patch-based) models.

Questions for authors

Please refer to the Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer cd58 (Part I)

Thank you for your constructive comments on the evidence for our main claim, experimental design, baseline comparisons, evaluation coverage, and mechanism analysis.

W1-1: The evidence for the main claim is insufficient. The paper claims that irregular forecasting does not require a specialized architecture and that general-purpose forecasting components are sufficient. However, the controlled study mainly tunes a single standard patch model.

A:

On the concern that the controlled study relies on a single standard patch model: BridgeTST was built step by step from a standard patch-based architecture commonly used for regular time-series forecasting. Patch-based models are widely used in regular forecasting, from PatchTST to recent foundation models such as TimesFM, Sundial, and Chronos2. We therefore chose a patch-based model for our controlled analysis.

On the concern that the process is simply step-by-step hyperparameter tuning: Our analysis uses the results from each stage to guide the next design choice. This is a progressive analysis, not just step-by-step hyperparameter tuning.

W1-2: The evaluation covers only 5+5 datasets, and the module-replacement experiments cover only three irregular datasets.

A:

Short-term regular datasets: We also evaluate short-term forecasting on two additional datasets. ILI uses an input length of 104 and forecast horizons of 24, 36, 48, and 60. FRED-MD uses an input length of 36 and the same forecast horizons. The table reports the average over the four horizons.

ModelILIFRED-MD
MetricMSE / MAEMSE / MAE
TFMixer3.9066 / 1.292666.1217 / 1.4614
APN4.9024 / 1.6181408.3594 / 4.2185
tPatchGNN5.6704 / 1.6936410.6707 / 4.2898
TimeMixer1.6079/0.7727‾\underline{1.6079 / 0.7727}68.6378 / 1.4348
iTransformer1.7663 / 0.816566.4492 / 1.3507
PatchTST1.9225 / 0.885461.6401/1.3344‾\underline{61.6401 / 1.3344}
Crossformer3.6768 / 1.2354399.9715 / 3.8326
DLinear2.0267 / 0.976885.1869 / 1.5322
BridgeTST1.5598 / 0.772451.8344 / 1.2091

The short-term setup is described in Section 5.1 of the revised paper‾\underline{\text{Section 5.1 of the revised paper}}. Per-horizon results and their analysis for ILI and FRED-MD are in Appendix A.6, Tables 11–12‾\underline{\text{Appendix A.6, Tables 11–12}}.

Coverage of irregular datasets: Representative studies on irregular forecasting evaluate their models on 4–5 irregular datasets. We evaluate BridgeTST on five commonly used irregular datasets.

Model / studyHuman ActivityUSHCNP12 (PhysioNet'12)MIMIC-IIIMIMIC-IV
GraFITi×✓✓✓✓
HyperIMTS✓✓✓✓✓
tPatchGNN✓✓✓✓×
APN✓✓✓✓×
TFMixer✓✓✓✓×

Coverage of module-replacement experiments: The staged component ablations on five irregular datasets are in Section 4, Tables 2–5‾\underline{\text{Section 4, Tables 2–5}}. They cover five datasets, not just three. We also report additional results for Electricity and Traffic from the regular-data module-replacement experiments in Section 5.3.1‾\underline{\text{Section 5.3.1}}. The results follow the same trend as the other three datasets.

DatasetBridgeTSTTTCNFTC HeadTTCN + FTC Head
MetricMSE / MAEMSE / MAEMSE / MAEMSE / MAE
Electricity0.1561 / 0.24830.1913 / 0.29000.8356 / 0.69180.7894 / 0.6678
Traffic0.4032 / 0.24710.6423 / 0.32491.2798 / 0.68731.2040 / 0.6329

Figure 3 in Section 5.3.1‾\underline{\text{Figure 3 in Section 5.3.1}} reports the module-replacement results on all five regular datasets.

W1-3&W2-1: The baseline settings are inconsistent, making the paper's conclusions too strong.

A:

For irregular forecasting models, we use the configurations provided by their public implementations for the corresponding irregular datasets. For settings without a public implementation, such as evaluating regular forecasting models on irregular datasets, we use the same hyperparameter search strategy as BridgeTST to make the comparison as fair as possible.

ModelsTraining objective
APN, GraFITi, HyperIMTS, tPatchGNNMSE
DLinear, PatchTST, iTransformer, TimeMixer, Crossformer, BridgeTSTMAE
TFMixerMAE + γ × reconstruction MAE

W2-2&W3-1: The use of RevIN is manually switched based on whether the data are regular or irregular.

A: RevIN is enabled or disabled based on directly observable sampling information. In our benchmark, regular data are uniformly and densely sampled, while irregular data have uneven intervals or missing observations. A simple data check can distinguish the two settings. No additional model or complex decision rule is needed.

AuthorsOpenReview ↗

Response to Reviewer cd58 (Part II)

W2-3: Some irregular forecasting baselines lack results on regular datasets.

A: To further evaluate GraFITi and HyperIMTS on regular time series, we evaluate them on ETTh1 and ETTm1 with an input length of 512 and a forecast horizon of 96. We compare them with BridgeTST.

DatasetModelTest resultRelative increase vs. BridgeTST
MetricMSE / MAEMSE / MAE
ETTh1BridgeTST0.3650 / 0.3892—
ETTh1GraFITi0.7587 / 0.6249+107.9% / +60.6%
ETTh1HyperIMTS0.8290 / 0.6576+127.1% / +69.0%
ETTm1BridgeTST0.2818 / 0.3281—
ETTm1GraFITi0.6489 / 0.5851+130.3% / +78.3%
ETTm1HyperIMTS0.7280 / 0.5931+158.4% / +80.8%

GraFITi and HyperIMTS have substantially higher forecast errors than BridgeTST on both regular datasets. Both methods build graphs or hypergraphs over individual observations. This is computationally expensive on dense regular time series. Limited compute and rebuttal time prevented us from evaluating them on the remaining regular datasets.

W3-2: The paper uses module-replacement results to explain why specialized modules hurt performance on regular data, but lacks fine-grained diagnosis or theory.

A: To better understand the performance degradation observed in Figure 3, we further conduct ablation studies on TTCN.

TTCN forms a patch representation by weighting each position's input with a position-specific filter coefficient, applying the same operation across output channels. For one output channel, let tt index a patch position; its input combines the observed value and timestamp encoding:

ut=[mtxt;ϕ(τt)].\mathbf u_t=[m_tx_t;\phi(\tau_t)].

Here, xtx_t is the observed value, mtm_t indicates whether the position is valid, and ϕ(τt)\phi(\tau_t) is the timestamp encoding. The original TTCN computes coefficients from the current input. We also introduce a static ablation in which the coefficients do not depend on the input. They are learned directly:

g_\theta(\mathbf u_t), & \text{TTCN},\\\\ \mathbf s_t, & \text{static TTCN}. \end{cases}$$ We also study how TTCN aggregates observations using these coefficients. The original TTCN uses softmax. We compare it with raw sum and mean sum, where $\mathcal V$ denotes the valid time positions: $$\mathbf w_t=\begin{cases} \operatorname{softmax}_{t\in\mathcal V}(\mathbf a_t), & \text{softmax},\\\\ \mathbf a_t, & \text{raw sum},\\\\ \mathbf a_t/\max(1,|\mathcal V|), & \text{mean sum}. \end{cases}$$ Softmax normalizes coefficients over valid positions. Raw sum uses the coefficients directly. Mean sum divides the sum by the number of valid positions. The weighted inputs are then summed to produce the output for that channel: $$z=\sum_{t\in\mathcal V}\mathbf w_t^\top\mathbf u_t.$$ We replace the input embedding in BridgeTST and compare five settings on ETTh1 and ETTm1 with an input length of 512 and a forecast horizon of 96: MLP, Original TTCN, Static TTCN with softmax, Static TTCN with raw sum, and Static TTCN with mean sum. | Input embedding | ETTh1 | ETTm1 | |---|---:|---:| | **Metric** | **MSE / MAE** | **MSE / MAE** | | MLP | **0.3650 / 0.3892** | **0.2818 / 0.3281** | | Original TTCN | 0.6830 / 0.5550 | 0.3745 / 0.4135 | | Static TTCN / softmax | 0.6504 / 0.5737 | 0.3332 / 0.3811 | | Static / raw sum | 0.4441 / 0.4585 | 0.3040 / 0.3488 | | Static / mean sum | 0.5974 / 0.5441 | 0.3442 / 0.3837 | Replacing softmax with raw sum substantially reduces the error, but the result is still worse than MLP. This suggests that softmax aggregation accounts for part of TTCN's performance drop. The revised manuscript therefore limits the mechanistic claims accordingly. TTCN's detailed ablation is in $\underline{\text{Section 5.3.2}}$, with results in $\underline{\text{Table 7}}$. **W3-3: Applying the four-stage framework to patch-based models that do not use canonical pre-alignment (CPA) may be an oversimplification.** **A:** Normalization, input embedding, dependency modeling, and forecast generation are common stages in time-series forecasting. Other papers use similar perspectives; TimeRecipe is one example (Zhao et al., 2026). We organize $\underline{\text{Section 3}}$ around these four stages and discuss the design and implementation of different methods at each stage. This structure provides a shared basis for comparison while preserving each method's architectural and implementation differences. It therefore does not oversimplify the analysis. **References** Zhao, Z., Ni, J., Xu, S., Liu, H., Jin, W., & Prakash, B. A. (2026). [TimeRecipe: A Time-Series Forecasting Recipe via Benchmarking Module Level Effectiveness](https://openreview.net/forum?id=CsoR8ztROC). *International Conference on Learning Representations (ICLR).*

Thanks for authors' detailed rebuttals. I have increased the informativeness and soundness scores. However, I still believe my initial overall rating reflects my assessment of the paper, and therefore I would like to keep the current rating.

AuthorsOpenReview ↗

Response to Reviewer cd58

Thank you for acknowledging that we have addressed your concerns. Please let us know if any issues remain. If you feel that our revisions have resolved them, we would appreciate your consideration of updating your overall score.

Comments on the paper

3 comments
~Junghoon_Lim1OpenReview ↗

Related Work: QuITE (ICML 2026)

Thank you for sharing this interesting work. I am an author of QuITE: Query-Based Irregular Time Series Embedding (ICML 2026), and I noticed some similarities in the motivation and key findings between this paper and our previous work.

In QuITE, we investigate whether existing regular time-series models can effectively handle irregular observations. We identify conventional input embedding mechanisms, rather than the backbone architectures themselves, as a key limitation, and demonstrate that regular forecasting models can be adapted to irregular time series through an appropriate embedding module without modifying their backbones.

This finding appears closely related to a central observation of BridgeTST: that general-purpose forecasting components can effectively handle irregular time series when appropriate input information is provided. QuITE may also provide useful context for the reviewers' concerns regarding whether the limitations of regular forecasting models stem from their backbones or input representations, as it directly investigates this question through embedding-based adaptation.

I also noticed that TFMixer (ICML 2026) is discussed and included in the experimental comparisons, whereas QuITE, published at the same conference, is not mentioned in the current manuscript. Given its relevance to the central motivation and findings of this work, I believe QuITE would be a valuable addition to the related-work discussion.

I recognize the distinct contributions of BridgeTST, particularly its systematic component-wise analysis and evaluation across both regular and irregular forecasting settings. A brief discussion of its relationship to QuITE would help clarify the shared insights and respective contributions of these works.

I hope the authors will find this reference useful.

Reference: QuITE: Query-Based Irregular Time Series Embedding, ICML 2026.

AuthorsOpenReview ↗

Response to QuITE Public Comment

Thank you for pointing out the connection between QuITE and our work. QuITE proposes a plug-in, query-based input embedding. It encodes irregular observations as tokens, then uses learnable queries and a single masked self-attention layer to aggregate them. This produces a representation compatible with existing multivariate time-series backbones. It requires neither interpolation nor changes to the backbone. QuITE also proposes QuITE++, a hierarchical forecasting architecture that combines patch-level and variable-level attention with a cross-attention decoder for future timestamps. The revised manuscript now includes QuITE in the discussion of query-based representations in Section 3‾\underline{\text{Section 3}}.

Both works study how to use information from irregular observations, but they have different scopes. QuITE focuses on input embedding, while BridgeTST studies the full forecasting pipeline. We analyze five methods with different architectures, covering both direct asynchronous modeling and canonical pre-alignment (CPA). We examine four stages: normalization, input embedding, dependency modeling, and forecast generation. In controlled experiments, we test the design choices at each stage. The results show that general-purpose forecasting components can form an effective forecasting pipeline when the input includes information about irregular sampling. We build BridgeTST and evaluate it on five irregular and five regular datasets. Building on these results, we further analyze the boundary between regular and irregular forecasting.

Dear Reviewers cd58, zYgH, and Kipt,

The authors have responded to your questions. Please review their responses and consider whether your concerns have been addressed. You may also take a look at the other reviewers' comments and the revised manuscript and share any further feedback.

Best regards,

Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.